Pull Action Items Out of a WhatsApp Group with Claude (2026)
A project group ran for a week and nobody wrote down who promised what. Ask Claude to read the group and list the commitments, deadlines and open questions, without sending anything.
Daniel Roth6 min read
On this page
Someone said they would send the tile quote. Someone else would call the electrician, and a third person asked whether the door was 80 or 90 centimeters and never got an answer. It is all in the group, somewhere between the photos and the thumbs-up emojis, and by Friday nobody remembers who owes what.
Ask Claude to read the group and pull out the commitments. With the Blueticks WhatsApp connector, Claude reads a group's recent messages, read-only, and lists who committed to what, the deadlines that were mentioned, and the questions nobody answered. It does not send, react or mark anything as read.
A group digest summarizes everything that happened. This prompt keeps only what someone is expected to do.
What does Claude read?
Claude calls the chats tool with the list_messages action. That action is marked read-only in the tool definition, so it is a read. It can take one group by chat_id, or up to 25 chats in one call with chat_ids.
The parameters that matter here:
| Parameter | What it does |
|---|---|
chat_id or chat_ids | One group, or up to 25 chats in a single call |
since / until | ISO 8601 bounds with a UTC offset, for example the last seven days |
limit | Messages per chat, up to 200. When you batch chats, it defaults to 20 |
query_any | Up to 24 keywords; a message matches if any one hits its text, caption or filename |
Claude does not need the chat ID up front. It runs a chats search on the group name and reads the ID from the result. If two groups have similar names, Claude should ask which one you mean.
Each message row carries what you need to attribute a commitment. author is the sender's ID, senderName is the display name when it can be resolved, timestamp is the send time, and text is the message text or caption. Messages you sent carry your own ID as author, so your own promises show up too. When no name resolves, you get a number, and the answer should say so.
The prompt: commitments, deadlines and open questions
Paste this into a normal Claude chat and change the group name:
Using Blueticks, read the last 7 days of messages in the WhatsApp group "Riverside Renovation Crew". Get today's date first. Find the group's chat ID with a chats search, then read its messages since that date with a limit of 200. From the messages, list three things: (1) commitments, meaning who said they would do what, and any date they gave; (2) deadlines or dates mentioned, with the message date; (3) questions asked in the group that nobody answered. For each item give the sender's name, or their number if no name is shown, and the message date. Paraphrase and do not quote more than a few words. Then tell me the date of the oldest message you read and whether the search was truncated. Do not send, react, mark as read or archive anything.
Claude makes about three calls. utils returns the current date, so "last 7 days" becomes a real timestamp. chats searches for the group. Then chats runs list_messages on that chat with since set. Claude reads the rows and the coverage note, then writes the list.
A sample answer
The group, names and dates below are invented to show the shape of the answer:
Commitments
- Dana: will send the revised tile quote (said 22 Oct, no date given)
- Lior: will book the electrician for next week
Deadlines
- Plumber deposit due 30 Oct (mentioned by Maya, 21 Oct)
Open questions
- Is the kitchen door 80 or 90 cm? Asked by Omer on 20 Oct, no reply
Oldest message read: 18 Oct. The scan was not truncated.
That last line tells you how far to trust the list.
What can go wrong?
- The list is not a record. A summary is Claude's reading of the messages. It can miss a promise buried in a long reply, or count a joke as a commitment. For anything important, open the group and check the source message before you act.
- The scan was truncated. The
coveragenote has atruncatedflag. When it istrue, the scan hit its cap and some messages may be missing. Shorten the date range, or name the chat and ask again. - Batches can skip chats. A batch has a time budget (
deadline_ms, 120000 ms by default). A chat that did not fit comes back marked as skipped, and it was not read. Claude should report it as not covered. Ask about skipped groups on their own. - The default limit is low. Batched chats default to 20 messages each. A busy group can have more than that in a week, so set
limitexplicitly, as the prompt does. - Keywords hide things.
query_anyonly returns messages that contain one of your words. "I'll sort it" will not match "quote" or "electrician". Use keywords to find a topic, and leave them out to see the whole window. - The engine is offline. Filtered reads do not fall back to the stored backup copy. If your engine is offline during a filtered search, you can get an empty page, and an empty page proves nothing. Reconnect and run it again.
- Free plan quota. The free plan allows 5 requests per 6-hour UTC slot, shared between the connector and the REST API. Each HTTP request counts once, so a multi-step search uses several of your five.
Is it safe to let Claude read a group?
Reading is not sending, but the same chats tool also holds actions that change things: sending, reacting, marking as read and archiving. You decide what runs. Approve each tool call yourself, check the action name before you click, and do not choose "Allow always." A list_messages call reads. A send_message_text call sends.
Run the prompt in a normal chat, not in Research mode. Anthropic's guide to custom connectors explains that in Research, Claude can call connector tools without asking (Anthropic's custom connector guide).
The line "do not send, react, mark as read or archive" only asks Claude to hold back. Your approval of each call is what actually stops a send. Everything Claude reads in the search is shared with it in that conversation, so keep the scan to groups you are comfortable sharing that way.
Related guides
To pull a file out of the same groups, see how to find a file someone sent you on WhatsApp. To set up the connector first, read connecting WhatsApp to Claude. The full list of tools is in the WhatsApp MCP tools guide. Plans and prices are on the pricing page, and you can start free.
Frequently asked questions
Can Claude find action items in a WhatsApp group? Yes. With the Blueticks connector, Claude reads a group's recent messages and lists the commitments, deadlines and open questions, with the sender and date for each. It only reads.
Will it reply to the group or mark messages as read? Not if you approve only the read calls. The prompt tells Claude not to act, but your approval of each tool call is what stops it.
Is the list reliable enough to act on without checking? Check it first. A summary can miss a promise in a long message or read a joke as a commitment. Open the source messages for anything important.
What if a group is missing from the answer? Check whether it was skipped by the batch time budget, or whether the search was truncated. Ask again for that group on its own.
